I am working on a project involving inference of causal direction from purely observational data, and not time series (given several assumptions, of course). I've been using the CauseEffectPairs database to validate my method, but the ground truth in these datasets is based on the obviousness of true causal direction (e.g. altitude causes air temperature, age causes income). However, causal inference would be useful in the real world if it could identify a true causal direction in cases where it is not intuitively obvious.
Let X and Y be collections of L observations of variables with dimension n and m, respectively. I want to evaluate my method on datasets with the following characteristics:
- There exists a non-confounded causal relationship between X and Y.
- The causal direction is not obvious. That is, reasonable mechanisms could be proposed both for the case where X causes Y and for the case where Y causes X, such that an experimental intervention would be necessary to choose between the two hypotheses.
- There exists experimental evidence that identifies the true causal direction (i.e. there's a ground truth).
The data can be continuous, discrete, categorical, mixed, etc.
For example, a dataset containing years of education and income at age 30 for some sample of US citizens would meet the first two criteria: 1) There is a clear correlation between education and income. 2) It is not clear whether education actually gives people the ability to earn more, or if people with the ability to earn more are predisposed to spend more years in education. This dataset would meet the third criterion if an experiment had been published that randomly assigned some people to pursue higher education and others not to do so and then observed their income at age 30. For obvious reasons, nobody has actually conducted such an experiment. So this would be a situation where observational causal inference would be useful, but I can't use this data to validate my method due to lack of ground truth.
The best answers would list some publically available data that could be used to create a dataset with the above criteria or describe a set of measurements that could be easily carried out that would produce such a dataset.